UMM-Reflection: interleaved RL teaches unified models self-correcting image generation
Yijia Fan · hf · 2026-09-29
UMM-Reflection applies reinforcement learning to complete reflection trajectories inside unified multimodal models that can both see and render images, enabling diagnose-render-diagnose loops.
- Sibling trajectories share one initial image so group-relative advantage compares reflection strategies; trajectory-level advantage updates both reflection tokens and flow-based revisions
- No external critic or inference-time verifier needed; SFT alone can't find high-success repair paths
- On BAGEL, improves GenEval by 12.05 points over SFT, transferring to unseen benchmarks: WISE (+10.97), OneIG-Bench (+3.48), T2I-CompBench++ (+4.63)
More from Multimodal
- Harvard-MIT UGG model unifies video generation, depth and camera pose in one diffusion model — burny_tech · 2026-11-14
- CapCut Web Ultra Offers a Week of Unlimited Seedream and GPT Image 2.5 Generation — JaynitMakwana · 2026-11-14
- Solv3R demo impresses: 3DGS reconstruction runs in about a minute — jonstephens85 · 2026-11-14
- Dreamina allegedly paid $82k in 'labor' for a 4-minute AI video — AIandDesign · 2026-11-14
- CapCut Adds Seedance 2.5 for 4K Video Generation in AI Lab and Video Studio — CurieuxExplorer · 2026-11-14
- BAAI's encoder-free VLM family EVE: NeurIPS spotlight, ICCV highlight, open-sourced — liuziwei7 · 2026-11-14